BottleCap AI, a Prague-based startup launched by Google and Meta alumni, has raised a $7.5m seed round to build more efficient foundational AI models.

The round was led by 20VC and included a host of high-profile angels like former Stripe CTO David Singleton, Canva cofounder Cliff Obrecht, Synthesia cofounder Steffen Tjerrild Hansen, Supercell CEO Ilkka Paananen, Lovable founder Anton Osika and French billionaire Xavier Niel.

The company was founded last year by cofounders Tomas Mikolov, previously a research scientist at Google, Meta and the CIIRC (Czech institute of informatics, robotics and cybernetics); Jarsolav Beck, a repeat entrepreneur and founder of best-selling VR game Beat Saber, which was acquired by Meta in 2019; and David Herel, an AI researcher at the CIIRC.

The compute needs of foundational AI models, particularly large-language models (LLMs), are quickly escalating, meaning model providers like OpenAI require huge amounts of capital and processing power.

“There is clearly a narrative of adding compute and hoping for the best, which is not a sustainable way of working,” Beck tells Sifted. “[Cofounder Tomas Mikolov]’s research has been very focused on efficiency and it was clear the future would be following that direction.”

“It makes sense, if you’re thinking of efficiency, to start from scratch. That’s why we’re building new foundational models with our own architecture.”

Efficiency gains

BottleCap is developing new algorithms and architectures that it says have efficiency at their core, unlike LLMs produced by competitors. “Other companies like Google obviously focus on efficiency, but they don’t have it as a fundamental principle,” says Beck.

For now, the startup isn’t sharing the details of its secret recipe, apart from that it is moving beyond the transformers architecture and techniques like mixture-of-experts, which are typical methods to build LLMs.

It is also too soon, says Beck, to publish data on the efficiency gains the startup has achieved so far. The objective is to improve the efficiency of LLMs by 100x, and early results published last year showed the company’s first algorithm could reduce costs of training by up to 50%.

Commercialising

In addition to developing AI models, BottleCap is setting up an “app development studio” to turn research outcomes into LLM-powered applications for consumers and enterprises. “Because we do everything in-house, for whatever use case we can make the best possible LLM,” says Beck.

Instead of selling its models, the startup will commercialise apps built on top of the technology.

BottleCap is announcing its first consumer app, Pulse, which aggregates and summarises information and news across chosen topics. Beck says future releases are more likely to focus on the B2B segment, with plans to eventually create “a new generation of AI tools” for businesses.

“The apps we’ll produce will finance the research centre,” he says. “Many companies raise money on the promise their models will be better but that’s not really sustainable. We think we should get revenue from day one.”

With the fresh funding, the startup plans to release a few more apps to kick-start commercialisation, and to scale operations by hiring more researchers and product developers.